COVID-19 Case & Death Prediction for Ethiopia

Problem

During the pandemic, Ethiopian health centers needed reliable forecasts of new COVID-19 cases and deaths to plan capacity. Local data was far messier than the clean series used in most published work: daily testing volume fluctuated heavily, and both new cases and deaths swung widely (a standard deviation of ~733 daily cases against a mean of ~634). Simple linear and ARIMA models — the common choice in earlier Ethiopian studies — could fit the past but generalized poorly to future days, weeks, and months.

Approach

Working from the Our World in Data dataset (738 daily records from March 2020 to March 2022, ~67 columns), we cleaned and reframed the problem for machine learning:

• Dropped near-empty and single-value features; filled the rest with zero-fill and scikit-learn KNN imputation.
• Engineered seasonality features after EDA showed clear seasonal structure — peaks in the colder months and troughs in the hottest (July), with new cases and deaths moving together.
• Encoded the vaccination signal, which showed a clear inverse relationship with new cases through 2022.
• Benchmarked ML regressors against the linear/ARIMA baselines from prior work.

Results

• Modeling seasonality and vaccination meaningfully improved forecasts over the day-count linear baselines used in earlier local studies.
• Produced multi-horizon forecasts (day/week/month) usable for precautionary planning rather than only fitting historical points.
• Work was co-authored as a research study at the Adama Science and Technology University AI facility.

Learnings

• Real-world epidemiological data needs heavy cleaning and imputation before any model is meaningful.
• Domain-driven features (season, vaccination) often beat more complex architectures on noisy, low-volume data.
• Forecasts are only useful when framed for the decision — capacity planning — not for leaderboard accuracy.